Detection Of Dangerous Goods In Subway Security Check Based On Deep Learning

Nowadays,subways play a very important role in public transportation,and the flow of people is also increasing,accompanied by increasing security pressure.However,the current subway security inspections mainly rely on humans to observe whether they carry dangerous items.This method requires a lot of human resources,and people are prone to fatigue during long-term observation,resulting in missed inspection of dangerous items and poor stability.It is difficult to meet the needs of the security inspection site,which makes a huge safety hazard in a crowded public place such as the subway.Therefore,in order to save costs and ensure public safety,through deep learning,automatic analysis by a computer to identify dangerous items and their locations in security inspection pictures is of great significance.Aiming at the real-time and accuracy requirements of subway security inspection,this article studies the setting of the default box of the SSD algorithm and the improvement of small target detection accuracy in the target detection algorithm to identify and locate dangerous items.The main research contents are as follows:(1)In view of the problem that the initial size and position of the default box need to be set in the algorithm,which increases the detection time,an adaptive method is proposed to adjust the initial default box to make it well match the target object.Using the idea of RPN in the two-stage algorithm,predicting a binary objectivity score to adjust the position and size of the default box to match the target object,and the recommendation process of candidate regions is simulated in two-stage algorithm,but it does not require a lot calculation of region proposal.Experiments show that this process provides a large number of balanced positive training samples during the training process,which improves the training speed.(2)In order to solve the mismatch between the receptive field of the prediction module and the object features suggested by the adjusted default box,which leads to the reduction of detection accuracy,an effective adaptive process is proposed to adjust the receptive field.The sampling points of the convolution filter are transformed according to the adjusted default box,and the offset of the adjusted default box is covered in an adaptive manner to increase the fixed grid sampling position of the convolution filter.Experiments show that this process brings a significant improvement in accuracy on various scale objects.(3)In order to improve the detection accuracy of small-sized target objects,an SSD algorithm based on feature fusion is proposed.Based on the SSD network,design enhancement modules for the shallow network to enhance the network’s extraction of small target features,and then fuse the features of the deep network target with the features of the shallow network target to obtain a rich level of fused features to achieve classification of various types of information at different network layers.Experiments show that this algorithm improves the detection accuracy and efficiency of small target objects.